RuntimeError: CUDA out of memory.

[Yi-Qi638]

I used the 1080Ti 11gb. but when I run the t.test(),sr = self.model(lr, idx_scale) I got the RuntimeError: CUDA out of memory. how can I fix this? thanks

[zcpchaos]

I also tried on a Tesla V100 GPU with 32Gb memory, but it still out of memory. I woder how did the owner conduct the experiments. Or something else is wrong?

[Magauiya]

I am also having the same issue with 2 GPUs (48 Gb Tesla). It is fascinating that the author could train the model on Titan Xp ?!

Even though I have tried to input 1 batch with 48x48x3. the memory issue remains.

[Magauiya]

Problem solved! The issue was in torchsummary, where I put 256x256 images. Anyway, the problem is in NLB, which consumes too much memory because of large matrix multiplication: f = torch.matmul(theta_x, phi_x)

[supratikbanerjee]

Problem solved! The issue was in torchsummary, where I put 256x256 images. Anyway, the problem is in NLB, which consumes too much memory because of large matrix multiplication: f = torch.matmul(theta_x, phi_x)

So any suggestion on how can I run the test code on Titan V?

[Magauiya]

Problem solved! The issue was in torchsummary, where I put 256x256 images. Anyway, the problem is in NLB, which consumes too much memory because of large matrix multiplication: f = torch.matmul(theta_x, phi_x)

So any suggestion on how can I run the test code on Titan V?

Change batch size to 1, it should work even if the image is 256x256x3 (my case). Otherwise, if you have several Titan V, you should implement model sharding.

[jiandandan001]

I have the same problem in the testing phase, even i use the '--chop' option. Do you know the reason? I use 1 1080 ti GPU. Thanks

[chenjiachengzzz]

I have the same problem in the testing phase, even i use the '--chop' option. Do you know the reason? I use 4 1080 ti GPU. Thanks

[chenjiachengzzz]

I have the same problem in the testing phase, even i use the '--chop' option. Do you know the reason? I use 1 1080 ti GPU. Thanks

did you fix this problem

[wanjh1024]

AlexHex7/Non-local_pytorch#9 Reference this issue, I solved this problem. This is because the code written by PyTorch 0.3.1 .In PyTorch> 0.4.0 version,volatile is deprecated.Wrap the test code into torch.no_grad().